Creating a tensor from array versus a numpy array

[tgkolda]

It seems that ttb.tensor only supports creating a tensor from a numpy array. Is there any reason it cannot create a tensor directly from a Python array?

Right now, the user has to type the following:

X = ttb.tensor(np.array([8, 9, 2, 9, 6, 1, 3, 5]), (2, 2, 2));

I'm suggesting instead that the following should be possible:

X = ttb.tensor([8, 9, 2, 9, 6, 1, 3, 5], (2, 2, 2));

Also, I'm not sure where the creation is clearly documented. I sort of inferred that it had to be a numpy array, but it's not clearly stated nor is there any example where you specify a simple tensor like this one.

[ntjohnson1]

I see the potential additional user friendliness to the later. Will share some thoughts but I'd be interested to hear Danny's thoughts on it.

A python list is effectively a list of pointers so it can be a collection of arbitrary types, and isn't contiguous in memory. Internally we'd have to convert to a numpy array, and make a decision about when to complain that the conversion isn't appropriately accurate. If we do make this move we can probably generate a utility to help for it, and we'd want to push that interface change out across the project, since right now I think for all numeric containers we've be requiring numpy arrays as the defacto python standard.

As for the documentation can you clarify what you were looking at?
help(ttb.tensor) vs https://pyttb.readthedocs.io/en/latest/tensor.html we do have types in the signature, and a very basic example where we provide a numpy array to construct a tensor. But I'm not sure where a new user looks first.

[tgkolda]

I was looking at the tutorials. I think most folks will start there, though they are in a sort of random order rather than structured like in the MATLAB help pages. I was looking here: https://pyttb.readthedocs.io/en/latest/tutorial/class_tensor.html.

[dmdunla]

Lists in Python are akin to cell arrays in MATLAB, and are used as collections of arbitrary objects. They do not have the full set of methods that numpy.ndarrays have and thus would require a lot more error checking in pyttb across all methods to support this. Thus, I do not think that such support would be helpful.

That being said, I think we should update the tutorials and documentation to provide examples using explicit data (rather than generators only) as @tgkolda suggests above.

[tgkolda]

FWIW, this might be interesting. It discusses how numpy approaches creating arrays:
https://stackoverflow.com/questions/40378427/numpy-formal-definition-of-array-like-objects

I would suggest we follow their lead if feasible.

[tgkolda]

This may also be interesting. We can potentially import from Numpy:

numpy/numpy-stubs#13

This could also get at the shape issue.

[dmdunla]

@tgkolda My take on this is that we should leverage NumPy as much as possible as a dependency rather than trying to replicate some of the core functionality separately in pyttb. I think with such a small dev team, this would be too much overhead to develop our own array_like input types rather than relying on users to use NumPy as the go-between for standard Python types and data.

I'm happy to discuss this more if you want to keep it as an open issue. Otherwise, I will close it and we can consider it if/when we move beyond NumPy as the only back-end for our code data types.